Python: [Bug]: AG-UI local approval execution does not receive function_invocation_kwargs / tool runtime context #8132

Description

When using AG-UI (AgentFrameworkAgent / run_agent_stream) with tools that require human approval (approval_mode="always_require"), approved tools are executed on the resume/Allow path before agent.run(...).

That local approval execution builds tool runtime kwargs only from AG-UI run_kwargs (typically session / tools / options). It does not carry the documented tool runtime context that applications pass via agent.run(..., function_invocation_kwargs={...}).

As a result, tools that correctly read runtime-only values from FunctionInvocationContext.kwargs (for example user_id) work during the normal in-agent.run tool loop, but on the AG-UI Approve/resume path those same keys are missing from ctx.kwargs.

Expected behavior

After a user approves a pending tool call over AG-UI, the local approved-tool execution should see the same tool runtime context (FunctionInvocationContext.kwargs) that the application associates with the agent run via function_invocation_kwargs (as documented for function tools / runtime context).

Actual behavior

On the AG-UI resume/Allow path, approved tools are invoked with custom_args derived from run_kwargs only. Documented runtime values such as user_id from function_invocation_kwargs are not present on ctx.kwargs for that execution.

Relevant code (current mainline layout)

• python/packages/ag-ui/agent_framework_ag_ui/_agent_run.py
  • run_agent_stream: resolves/executes approval responses before calling agent.run(...)
  • _resolve_approval_responses: builds tool_kwargs from run_kwargs (excluding options) and passes them as custom_args into _try_execute_function_call_groups

Contrast with documented Agent API

@tool
def get_weather(location: str, ctx: FunctionInvocationContext) -> str:
    user_id = ctx.kwargs.get("user_id", "unknown")
    ...

await agent.run(
    "What is the weather like in Amsterdam?",
    function_invocation_kwargs={"user_id": "user_123"},
)

In a normal Agent.run tool loop, user_id is available on ctx.kwargs. On the AG-UI Approve/resume local-execution path, the equivalent runtime context is not available to the same tool.

Code Sample

Minimal shape (conceptual):

1. Expose an AG-UI agent with a tool that uses ctx: FunctionInvocationContext and reads ctx.kwargs["user_id"] (or any key supplied only via function_invocation_kwargs).
2. Drive a turn that hits approval_mode="always_require" so the run finishes with a pending approval / interrupt.
3. Resume/Allow the pending approval through AG-UI.
4. Observe that during local approved-tool execution, ctx.kwargs does not contain the runtime key that function_invocation_kwargs is documented to provide for tool invocation.

Error Messages / Stack Traces

No framework crash is required. The failure mode is missing runtime context on FunctionInvocationContext.kwargs during AG-UI approved-tool execution (application-level permission / identity checks then fail or fall back incorrectly).

Package Versions

agent-framework-core: 1.15.0
agent-framework-ag-ui: 1.2.0

Python Version

Python 3.14.2

Additional Context

This is specific to the AG-UI approval / resume local-execution path, not to plain Agent.run HITL handling that stays inside a single agent.run tool loop.